You are on page 1of 15

Journal of Hydrology 519 (2014) 3163–3177

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Comparing statistically downscaled simulations of Indian monsoon


at different spatial resolutions
K. Shashikanth a,b, C.G. Madhusoodhanan a, Subimal Ghosh a,c,⇑, T.I. Eldho a,c, K. Rajendran d,
Raghu Murtugudde c,e
a
Department of Civil Engineering, Indian Institute of Technology Bombay, Mumbai 400076, India
b
Department of Civil Engineering, University College of Engineering, Osmania University, Hyderabad 500007, India
c
Interdisciplinary Programme in Climate Studies, Indian Institute of Technology, Bombay, Mumbai 400076, India
d
CSIR Fourth Paradigm Institute (CSIR-4PI), Bangalore 560037, India
e
Department of Atmospheric and Oceanic Science, Earth System Science Interdisciplinary Center, University of Maryland College Park, MD 20742, USA

a r t i c l e i n f o s u m m a r y

Article history: Impacts of climate change are typically assessed with fairly coarse resolution General Circulation Models
Received 2 June 2014 (GCMs), which are unable to resolve local scale features that are critical to precipitation variability. GCM
Received in revised form 15 October 2014 simulations must be downscaled to finer resolutions, through statistical or dynamic modelling for further
Accepted 16 October 2014
use in hydrologic analysis. In this study, we use a linear regression based statistical downscaling method
Available online 24 October 2014
This manuscript was handled by G. Syme,
for obtaining monthly Indian Summer Monsoon Rainfall (ISMR) projections at multiple spatial
Editor-in-Chief resolutions, viz., 0.05°, 0.25° and 0.50°, and compare them. We use 19 GCMs of Coupled Model Intercom-
parison Project Phase 5 (CMIP5) suite and combine them with multi model averaging and Bayesian model
Keywords:
averaging. We find spatially non-uniform changes in projections at all resolutions for both combinations
Statistical downscaling of projections. Our results show that the changes in the mean for future time periods (2020s, 2050s, and
Bayesian multimodel average 2080s) at different resolutions, viz., 0.05°, 0.25° and 0.5°, obtained with both Multi-Model Average (MMA)
Uncertainty and Bayesian Multi-Model Average (BMA) are comparable. We also find that the model uncertainty
Signal to Noise Ratio (SNR) decreases with projection times into the future for all resolutions. We compute Signal to Noise Ratio
(SNR), which represents the climate change signal in simulations with respect to the noise arising from
multi-model uncertainty. This appears to be almost similar at different resolutions. The present study
highlight that, a mere increase in resolution by a way of computationally more expensive statistical
downscaling does not necessarily contribute towards improving the signal strength. Denser data
networks and finer resolution GCMs may be essential for producing usable rainfall and hydrologic
information at finer resolutions in the context of statistical downscaling.
Ó 2014 Elsevier B.V. All rights reserved.

1. Introduction system to increasing the green house gas concentrations


(Prudhomme et al., 2002; IPCC, 2007). Since GCMs typically oper-
India receives more than 80% of its total rainfall during the four ate at coarse resolutions, they often fail to capture the finer scale
summer monsoon months; June to September (Webster et al., processes that affect regional climate. Hence, the outputs from
1998; Menon et al., 2013). Agriculture, water supply, industrial GCMs cannot be directly used in the regional impact assessment
production and allied activities depend on the Indian summer studies, specifically in hydrologic applications (Salvi et al., 2013).
monsoon rainfall (ISMR hereafter) and hence reliable regional The new generation climate models in the CMIP5 (Coupled
projection of ISMR, under climate change is necessary for adapta- Model Intercomparison Project Phase 5) suite have undergone sev-
tion and policy framing. Impacts of climate change are assessed eral improvements in terms of physics and resolutions (Taylor
by General Circulation Models (GCMs), which are at present the et al., 2012). Still, outputs from CMIP5 models cannot be directly
most reliable tools to simulate the response of the global climate used for regional scale projections because of the lack of regional
specificity in their simulations due to the relatively coarse resolu-
tions. Such coarse resolution projections must be downscaled to
⇑ Corresponding author at: Department of Civil Engineering, Indian Institute of
obtain regional/local scale projections. They can be obtained either
Technology Bombay, Powai, Mumbai 400 076, India. Tel.: +91 22 2576 7319, fax:
+91 22 2576 7302. by statistical or dynamic downscaling approaches. In dynamical
E-mail address: subimal@civil.iitb.ac.in (S. Ghosh). downscaling, a higher resolution regional climate model (e.g.,

http://dx.doi.org/10.1016/j.jhydrol.2014.10.042
0022-1694/Ó 2014 Elsevier B.V. All rights reserved.
3164 K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177

RegCM, Limited Area Model (LAM), PRUDUNCE) is embedded approaches to extract regional renditions of future projections
within a GCM, which provides large scale boundary conditions. from these models. This is the main motivation of our study.
Such regional models also incorporate sub-grid/local scale features The uncertainty stemming from various formulations can be
to deliver high resolution outputs (Fowler et al., 2007; Maraun assessed by ensembles of simulations based on different GCMs,
et al., 2010). The main limitation of dynamic downscaling method or ensembles of particular GCMs with different initial conditions
is the requirement of huge computational resources (Wilby and or with perturbed physics. The important roles played by these
Wigley, 1997). In addition to the limitations of the models them- sources of uncertainty depend on the time scales. In the decades
selves in capturing the regional and local land, atmosphere, and of the immediate future, the projections are argued to be domi-
ocean interactions (see for e.g., Liang et al., 2012). nated by model uncertainties and internal variabilities whereas
Statistical downscaling belongs to the class of data driven in the following decades, scenario uncertainties are significantly
approaches, which not only overcomes the limitation of computa- larger (Hawkins and Sutton, 2011; Yip et al., 2011; Deser et al.,
tional requirement but also can be applied to any region of inter- 2012). However, the internal variability increases at smaller spatial
est given sufficient data for constructing the statistical model. scales and shorter time scales (Hawkins and Sutton, 2009).
Statistical methods are thus predominantly used when sufficient Prudhomme and Davies (2009) reported that uncertainties due to
lengths of historical data are available for establishing the requi- downscaling methods and emission scenarios are of similar magni-
site stable statistical relationships (Wilby et al., 2004). The philos- tude, and generally smaller than GCM uncertainty for the 2080
ophy of statistical downscaling method is based on the time horizon. The uncertainties associated with the use of multiple
development of a relationship between predictors (coarse resolu- models have been studied by many researchers (Ghosh and
tion climate variables/synoptic scale circulation patterns) and a Mujumdar, 2009; Watanabe et al., 2012; Woldemeskel et al.,
predictand (e.g., rainfall). They are classified into weather gener- 2012). Uncertainty across different GCMs has been measured by
ators (Wilks, 1999a,b; Khalili et al., 2009) weather typing/classifi- a metric/skill known as coefficient of variation (CV) (Watanabe
cation (Corte-Real et al., 1995) and transfer functions (Ghosh and et al., 2012; Shashikanth et al., 2013) where higher CV denotes high
Mujumdar, 2009). Statistical downscaling methods build different uncertainty. The CV is defined as the ratio of standard deviation to
types of empirical relationships (like parametric/nonparametric the mean and it is a dimensionless quantity. Shashikanth et al.
and/or linear/nonlinear) between large-scale atmospheric predic- (2013) measured the uncertainty resulting from the usage of
tor variables and the regional climate variable(s) of interest (pre- multiple GCMs from both CMIP3 and CMIP5 suites for ISMR simu-
dictand) (Wilby et al., 2004; Kannan and Ghosh, 2013). In lations for the future time horizon (2081–2100) and reported that
statistical downscaling, the regional climate is assumed to be con- statistically downscaled simulations of CMIP5 show lower uncer-
ditioned by the large-scale climate predictors/variables and regio- tainty compared to those of CMIP3.
nal/local physiographic and region-specific processes (Von Storch, The projections of climate variables are now presented in a
1995, 1999; Kannan and Ghosh, 2013; Dobler and Ahrens, 2011). probabilistic framework (Raftery et al., 2005). This has led to the
The statistical downscaling methodology has been extensively development of new methods like Bayesian Multi-model Averag-
used for regional applications all over the globe. Complete and ing (BMA) (Tebaldi et al., 2004, 2005) which is a better alternative
comprehensive reviews of downscaling techniques, their advanta- to Multi-Model Averaging (MMA), where equal weights are given
ges and limitations, recent developments can be found in the to all GCMs for both past and future projections. In the BMA, differ-
Maraun et al. (2010), Fowler et al. (2007), Schoof (2013). The var- ential weights are assigned to the GCMs, based on performances
ious statistical downscaling approaches are summarized in during the historic period and convergence (agreement with
Table 1. majority of ensemble members in the projection period). Miao
The results of regional projections are always prone to uncer- et al. (2013, 2014) applied the BMA technique for temperature sim-
tainties stemming from multiple sources. The main sources of ulations in the different land regions of globe and in the Northern
uncertainty are model uncertainty, uncertainty from anthropo- Eurasia and showed that BMA technique is superior to MMA.
genic/scenario forcing factors and natural variability (Hawkins Giorgi and Mearns (2002) have developed REA (Reliability
and Sutton, 2009; Maraun et al., 2010; Woldemeskel et al., 2012; Ensemble Average) method for estimating the probability of regio-
Yip et al., 2011). The use of multiple downscaling models may also nal climate change exceeding a given threshold based on ensem-
result in downscaling uncertainty (Dibike and Coulibaly, 2005; bles of different AOGCM model simulations. The method takes
Ghosh and Katkar, 2012; King et al., 2012). Model uncertainty is into account two reliability criteria as mentioned above; model
the main source of uncertainty and mainly attributed to incom- performance for the present climate and convergence of the simu-
plete understanding of geophysical processes and limitations in lated future changes by all models (convergence criterion). Tebaldi
solving numerical schemes, parameterizations, and resolutions et al. (2004, 2005) developed a Bayesian Multi-Model Averaging
(Stainforth et al., 2007; Räisänen, 2007). Anthropogenic forcings (BMA) approach to determine Probability Density Function (PDF)
and scenarios describe how plausible trajectories of different of temperature change from the outputs multi-model ensembles.
aspects of future climate includes emissions, socio-economic Duan and Philips (2010) applied the BMA technique for the esti-
assumptions, energy use, population, air pollutants, land use and mation of weights assigned to individual historical simulations and
consequent radiative forcing and temperature anomalies (Wayne, future climate projections. In their study, the authors used 17
2013). The internal uncertainty is due to chaotic climate system GCMs from the CMIP3 suite. They also computed the climate
and natural forced variability is in fact considered ‘unforced’ vari- change signal from multimodel projections and the Signal to Noise
ability since the only source of energy is the solar radiation Ratio (SNR), where noise refers to the uncertainty resulting from
(Maraun et al., 2010). It is also clear now that higher latitudes the disagreement between model simulations. Accurate regional
are dominated by natural variability and the anthropogenic cli- projections are argued to be a key research challenge in climate
mate signal may become distinctly identifiable only in the tropics change science (Schiermeier, 2010), and in the recent decade,
over the next several decades (Deser et al., 2012). Even in the efforts have been made for finer resolution projections. For
tropics, the relatively small decorrelation scales of precipitation dynamic downscaling, the key challenge lies in huge computa-
compared to temperatures offer severe challenges to the state-of- tional requirements, while for statistical downscaling, the require-
the-art GCMs in delivering reliable projections of rainfall ment is the availability of finest resolution data. With the
variability. It is thus important to understand the regional and local improvements in computational power, as well as with the avail-
uncertainties in these projections by combining statistical ability of newer fine resolution gridded products, the resolution
K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177 3165

Table 1
Statistical downscaling approaches to different case studies.

S. No. Literature citation Case study Features of the model


1 Bardaossy and River Ruhr (Germany) A multidimensional stochastic model is developed for the space–time distribution of daily
Plate (1992) precipitation
2 Hewitson and Tropical Mexico region Empirical relationships between sea level and 500 mb circulation fields, and the local
Crane (1992) precipitation in Chiapas, Mexico, are derived using an Artificial Neural Networks (ANN)
3 Corte-Real et al., Portugal The multivariate adaptive regression splines (MARS) model was used to capture the observed
1995 relationships between sea level pressure (SLP) anomalies over the Euro-Atlantic sector and
the winter time (December–February) monthly rainfall
4 Convoy and Jones, British Isles Analysis using Lamb weather types, counts of weather fronts and air flow indices show strong
1998 relationships with daily rainfall characteristics
5 Wilks(1999) 30 stations located in the A stochastic weather generator model is developed and used for downscaling
state of Oregon in US
6 Mehrotra and Sydney Australia K-nearest neighbor based nonparametric, non-homogeneous hidden Markov model is
Sharma (2005, developed to generate rainfall sequences at multiple locations using synoptic scale circulation
2006) patterns
7 Ghosh and Orissa, India A linear regression developed between GCM simulated predictors Vs rainfall data. Principal
Mujumdar (2006) component analysis (PCA) is used to reduce the dimensionality and fuzzy clustering
algorithm is used to classify the components
8 Anandhi et al. Malaprabha reservoir The monthly precipitation is downscaled using support vector machine
(2008) catchment in India
9 Raje and Mahanadi catchment Developed conditional random field based algorithm to downscale rainfall. Direct
Mujumdar (2009) Orissa, India classification of dry/wet days as well as precipitation amount is achieved within a single
modelling framework
10 Mehrotra and Network of 45 rain gauges Used stochastic downscaling framework for simulation of multisite daily rainfall occurrences.
Sharma (2010) near Sydney, Australia Rainfall occurrences are modelled using a modified Markov mode and rainfall amounts are
estimated using nonparametric kernel density simulator conditioned on previous time step
rainfall using large scale predictors
11 Ghosh (2010) Assam and Meghalaya, SVM based relationship between predictors and predictand. Probabilistic Global search
India algorithm, a probabilistic search technique is used to compute the optimum number of
parameters for SVM; uncertainty addressed using modified Reliability Ensemble Average
(REA) technique
12 Groppelli et al. Oglio River basin in Italy Stochastic space random cascade (SSRC) approach is developed to downscale rainfall from
(2011) GCM data under climate change scenario
13 Kannan and Ghosh Mahanadi River Basin, Developed the Kernel Regression based downscaling model conditioned on weather state
(2013) Orissa, India based on Classification And Regression Tree (CART) method
14 Salvi et al. (2013) India Used CART and Kernel Regression technique to project rainfall using CCCma GCM
15 Gutmann et al. USA The downscaling algorithms considered are daily and monthly-disaggregated-to-daily Bias
(2014) and Bias Corrected Constructed Analog (BCCA), Asynchronous Regression (AR), methods

for regional modelling is getting enhanced; however, the science the future to project the regional climate scenarios. Here, we use
question, which remains unaddressed, is whether there is a corre- the National Centre for Environmental Prediction/National Centre
sponding improvement in the skills of simulations with the for Atmospheric Research (NCEP/NCAR) reanalysis data as predic-
improvements in resolution. It is also important to understand tors and the gridded rainfall as the predictand. For calibration of
the implications of using higher resolutions on the model uncer- the model, we use the gridded rainfall data, provided by the Asian
tainty and specifically, how the SNR changes with the change in Precipitation Highly Resolved Observational Data Integration
resolution. The improvements in ensemble simulations with BMA towards Evaluation of Water Resources (APHRODITE) at multiple
when compared with that of Multi-Model Averaging (MMA) may resolutions.
also change with the change in resolutions and such evaluations The selection of predictors is a crucial step in developing a
have not been reported in literature. Here, in this present study, statistical downscaling model. The output of the statistical down-
we attempt to address these questions with the case study applica- scaling method is sensitive to the choice of predictor variables.
tion to the Indian Summer Monsoon Rainfall (ISMR). We apply sta- The criteria used for selections of predictors are that they should
tistical downscaling, at multiple resolutions, viz., 0.05°, 0.25° and be reliably simulated by GCMs, available in the GCM archive and
0.50 for entire India. We use 19 GCMs from the CMIP5 suite, which physically associated with the variable of interest (Wilby et al.,
is a fairly large number of models to project rainfall. The perfor- 2004). Considering these criteria, the predictors used for ISMR
mances of the models are evaluated at multiple scales, as well as downscaling (also used in Salvi et al., 2013; Kannan and Ghosh,
with a common scale, where the downscaling is followed by 2013; Shashikanth et al., 2013) are air temperature, wind velocities
re-gridding/interpolation. PDF based skill scores are used for eval- (U and V wind), specific humidity at both 500 hPa pressure level
uation (Perkins et al., 2007) of models. Both MMA and BMA are and the surface and the Mean Sea Level Pressure (MSLP). These
computed and compared at different resolutions. We also evaluate predictors are based on the study by Shashikanth et al. (2013) for
the models for their SNR (Duan and Philips, 2010) at different res- ISMR.
olution to understand the impacts of resolution on SNR. The next The NCEP/NCAR reanalysis data (Kalnay et al., 1996), used for
section presents the details of the data used for this study. the predictors, are extracted for latitudes 5°N–40°N and longitude
60°E–120°E, encompassing the entire India. This is a global gridded
2. Data product and continually updated through data assimilation to judi-
ciously blend physical observations and model simulations dating
A statistical downscaling model (Fig. 1) first develops an empir- back to 1948. The Reanalysis data is a surrogate for observed data
ical relationship between synoptic scale circulation patterns (pre- for any predictor variable. The resolution of NCEP/NCAR reanalysis
dictors) and the local variable of interest (predictand). The product is 2.5° lat  2.5° long. For downscaling purposes, India is
relationship is then further applied to the GCM simulations of first subdivided into seven meteorological subdivisions as per India
3166 K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177

3.1. Statistical downscaling

The statistical relationship, developed between the reanalysis


predictors and gridded rainfall is based on a linear regression, sim-
ilar to the ‘‘Statistical DownScaling Model’’ (SDSM), developed by
Wilby and Dawson (2004). The methodology is presented in
Fig. 1 and is applied to the seven meteorologically homogeneous
zones (Parthasarathy et al., 1996) over India. The gridded
multi-dimensional predictors are highly correlated amongst each
other and together form high dimensional covariates. This leads
to multicollinearity in a linear regression. As the present model
is developed for monthly monsoon rainfall and the sample size
of calibration set length is small, and the model may over fit the
data with high dimensional predictors. Hence, we apply Principal
Component Analysis (PCA) to reduce dimensionality by retaining
the major fraction of the variability of the predictor field. Here,
we use first few principal components (PCs) which together repre-
sent 90% of the variability of original predictors. The principal com-
ponents (PCs) thus obtained are regressed with the rainfall, at
individual grids. The set of predictors and hence the principal com-
ponents, used as regressors, remain unchanged for all the grids in a
homogeneous region. The model is calibrated for the period 1970–
2005, for all three spatial resolutions. Applying the principal direc-
tions obtained from reanalysis for computing the PCs, to the GCM
output may introduce new bias. However, this is a standard proce-
dure (Wilby et al., 2004), which is followed for the GCM outputs, to
apply the same relationship, which is obtained between reanalysis
predictors and observed predictand. Obtaining the individual prin-
cipal direction from each GCM will not ensure this and hence is not
followed. A possible solution would be merging of reanalysis and
GCM predictors for obtaining the PCs, and is a potential area of
Fig. 1. Statistical downscaling (SD) algorithm used for the study. future research.
The calibrated model is then applied to the GCM simulations
after bias correction. GCM simulations are first regridded to the
Meteorological Department (IMD) categorization (Parthasarathy reanalysis grids, and then Wilby et al. (2004) method of standard-
et al., 1996) and the spatial extent of the predictors for different ization is applied. The principal directions obtained from the
zones are different (Fig. 2). They are selected based on the correla- reanalysis data is applied to the bias corrected GCM predictors
tion between the spatially averaged rainfall and gridded predictors for obtaining the principal components (PCs) corresponding to
(Salvi et al., 2013). GCMs. These PC are then used in the calibrated linear regression
The gridded rainfall data, provided by APHRODITE (Yatagai model to obtain gridded rainfall simulations for the historic (here,
et al., 2012) is used as the predictand, and is obtained at monthly we consider, 1970–2005) and future (2010–2099) periods (RCP 4.5
intervals for the spatial resolution 0.5°, 0.25° and 0.05°. They are scenario).
based on (i) GTS (Global Telecommunication system) data (ii) data
compiled by the organizations from the respective countries like,
India Meteorological Department (IMD) for India, and (iii) APHRO- 3.2. Evaluating the GCM simulations
DITE’s own data collection system with proper quality control
(Yatagai et al., 2012). For historic and future simulations, we The downscaled GCM simulations are first evaluated based on
selected 19 GCMs, from CMIP5 archives, and obtained the outputs the deviations from the long term mean and standard deviations
at monthly time scale. The list of GCMs is provided in Table 2. The for the historic period. However, such evaluations may not com-
GCM simulated predictors are obtained from the Program for Cli- pare the observed and projected simulations in terms of all the
mate Model Diagnosis and Intercomparison (PCMDI). For future statistical characteristics and variability. Hence, here we apply
rainfall projections of ISMR, Representative Concentration Path- the evaluation metric, suggested by Perkins et al. (2007), where
ways 4.5 (RCP4.5) is considered for the present work. RCP 4.5 rep- the skill of the simulation is measured with respect to the
resents a stabilization scenario where the total radiative forcing is resemblance of its Probability Density Functions (PDFs) to that
stabilized at 4.5 W/m2 at 2100, by employing a range of technolo- of observed.
gies and strategies to reduce greenhouse gas emissions (Wise et al., In this evaluation method, first the entire range of observed/
2009). simulations are divided into multiple bins. This methodology cal-
culates the cumulative minimum value between the model and
observed data distribution in each of bin; thereby measuring the
3. Methodology
common area between two PDFs. If a model simulates the
observed data perfectly, the skill score is equal to 1. If it simulates
In the present analysis, we first downscale the monthly rainfall
the data poorly, the skill score is equal to zero. Mathematically, the
to 0.50°, 0.25° and 0.05° spatial resolutions, with 19 GCMs. The
Skill score (S-score is given by)
resulting uncertainty is modelled with Bayesian modelling and
the SNR is evaluated for each of the resolution. Brief overview of X
n

downscaling, uncertainty modelling and estimation of SNR are S-score ¼ minimumðZ m ; Z o Þ ð1Þ
i¼1
discussed in the following subsections.
K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177 3167

(A)

Jammu and Kashmir Zone


Lat. 27.5N - 40 N
Long. 70 E - 82.5 E
NEH Zone
Lat. 25 N – 32.5 N
Long. 90 E - 100 E
Western Zone North Zone
Lat 17.5 N - 35N Lat. 15 N – 32..5 N (B)
Long 65 E - 80 E Long. 72.5 E – 92.5E
Central Zone
Lat. 12.5 N - 30 N
Long. 70 E - 87.5 E

North East Zone


Lat. 20 N – 32.5 N
Long. 85 E - 100 E

South Zone
Lat. 5 N – 22.5 N
Long. 70 E - 90 E

Seven Meteorologically homogeneous zones of


Indian Region as per India Meteorological
Department (IMD), (Parthasarathy et al.,1996)
with spatial extent of predictors

Fig. 2. Meteorological homogeneous regions of India identified by IMD (A) (Parthasarathy et al., 1996) and the areal extent of predictors for the central zone (B).

Table 2
List of GCMs used in monthly rainfall downscaling.

S. no. Name Institution Resolution


Lat (°) Long (°)
1 BCC-CSM1 Beijing Climate Centre 2.7906 2.8125
2 BNU-ESM College of Global Change and Earth System Science, Beijing Normal University 2.79 2.81
3 CCCma-CAN-ESM2 Canadian Centre for Climate Modelling and Analysis 2.8 2.8
4 CMCC-CMS Centro Euro-Mediterraneo per I CambiamentiClimatici 1.87 1.88
5 CNRM-CM5 Centre National de Recherches Meteorologiques 1.4 1.4
6 CSIR-ACCESS 1–0 CSIRO (Commonwealth Scientific and Industrial Research Organization, Australia), 1.25 1.875
and BOM (Bureau of Meteorology, Australia)
7 CSIRO-Mk3–6-0 Commonwealth Scientific and Industrial Research Organization in collaboration 1.87 1.88
with Queensland Climate Change Centre of Excellence
8 FIO-ESM The First Institute of Oceanography, SOA, China 2.79 2.81
9 GISS-E2-R-CC NASA Goddard Institute for Space Studies 2.00 2.5
10 Had-GEM2-E2S Met Office Hadley Centre 1.875 1.25
11 INM-CM4 Inst. for Numerical Mathematics 1.5 2
12 IPSL-CM5A-LR Institute Pierre-Simon Laplace 1.89 3.75
13 IPSL-CM5A-MR Institute Pierre-Simon Laplace 1.89 3.75
14 MIROC5 Atmosphere and Ocean Research Institute (University of Tokyo)and others 1.4 1.41
15 MIROC-ESM Atmosphere and Ocean Research Institute (University of Tokyo)and others 2.8 2.8
16 MPI-ESM-MR Max-Planck-Inst. For Meteorology 1.8652 1.875
17 NASA-GISS-E2-R NASA Goddard Institute for Space Studies 2 2.5
18 NorESM1-ME Norwegian Climate Centre 1.8947 2.5
19 NorESM1-M Norwegian Climate Centre 1.5 1.9
3168 K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177

where n = total number of bins; Zm = frequency values in a given bin where (i) and s are iterations and spatial points, respectively, S is the
from model m; Zo = frequency values in a given bin from observed total number of spatial points. For further details refer to Duan and
data. Philips (2010) and Raftery et al. (2005).
The S-score is computed for each grid, and then the spatial dis-
tribution of this score is obtained in terms of the PDF. The PDFs of 3.4.1. BMA consensus Expectation and uncertainty estimation
S-scores, obtained from multiple resolutions are compared. The BMA expected mean of posterior PDFs for the past and
future climates are estimated as a weighted sum of individual
3.3. Uncertainty modelling with Bayesian approach model simulations and are calculated as below:
X X
EðY P jf p1 ; f p2 ; . . . ; f pk Þ ¼ wk  EðY P jf pk Þ ¼ wk  f pk ð6Þ
Conventional approach for simulating projections with multiple k
models is to obtain the multi-model average, assuming that the
X X
positive and negative errors in the models will cancel each other EðY F jf F1 ; f F2 ; . . . ; f Fk Þ ¼ wk  EðY F jf Fk Þ ¼ wk  f Fk ð7Þ
out (Knutti, 2010), resulting in more reliable projections. Following k
Duan and Philips (2010), here we briefly discuss the steps for
assignment of weights to GCMs, with the Bayesian Model Averag- The total variances are the measures of spreads (uncertainties)
  of the posterior PDFs and are given by the following expressions:
ing (BMA). The parameters to be determined are h ¼ wk ; r2PK ; r2FK ,
where wk are the weights to each GCM (wk for kth GCM), rPK are 2 !2
X X X
the variance for each of the past/historical downscaled GCMs and r2 ðY P jf p1 ; f p2 ; . . . ; f pk Þ ¼ wk  f PK  wk  f PK
r2FK are the variance for each of downscaled GCM’s future k k k
X
projections. þ wk  r2 ðY TP jf PK Þ ð8Þ
k
Step 1: Let us assume the vectors y = (YP, YF) denote the down-
scaled data set consisting of past-climate (YP) and sim- !2
X X X
ulated climate variable for the future (YF) from each of 2
r ðY F jf F1 ; f F2 ; . . . ; f FK Þ ¼ wk  f FK  wk  f FK
the GCMs. The two data sets are assumed to be inde- k k k
X
pendent of each other and follow Gaussian þ wk  r2 ðY TF jf FK Þ ð9Þ
distributions. k

gðyjhk Þ ¼ gðyp ; yF jhPk ; hFk Þ ¼ gðyP jhPk Þ  gðyF jhFk Þ ð2Þ

where hPk, hFk denote parameters for past and future downscaled 3.5. Estimation of Signal to Noise Ratio (SNR)
simulated GCM output. The above equation is based on Bayes’
theorem. The estimation of signal to noise ratio is the measure of the
Step 2: The log likelihood function for Eq. (2), becomes robustness of a projection of a climate variable (rainfall)
X X (Hawkins and Sutton, 2011). This is estimated with the help of
lðhÞ ¼ log wk  gðY TP jhPK Þ  gðY TF jhFK Þ ð3Þ downscaled climate variables for the past and the future in individ-
ðsÞ k ual models ({fP1, fP2,. . ., fPK}, {fF1, fF2,. . ., fFK}) from GCMs along with
where YTP and YTF are evidentiary targets (observed) and weighted the BMA parameters h ¼ fW k ; r2PK ; r2FK g. The steps are as follows
multimodel consensus estimates of climate variable for the future, (Duan and Philips, 2010);
respectively.
Since g(YP|hP) and g(YF|hF) are Gaussian functions, the maximum i. Select the ensemble size of 1000. Set spatial point i = 1, and
likelihood estimates of the means lk of the individual model ensemble j = 1.
simulations are equal to the mean of the simulations them- ii. A GCM is selected randomly based on uniform random num-
selves; i.e., lPk = fPk and lFK = fFk. The fPk and fFk are the means ber and a probability is generated randomly. The values are
of downscaled simulations of the climate variable for the past obtained as YPK(i,j) from the PDF of normal distribution
ðiÞ ðiÞ
and future in each GCM. gðY PK jf PK ; r2PK Þ and YFK(i,j) from the PDF of normal distribu-
Step 3: The parameters h = {wk, r2PK ; r2FK } in Eq. (3) are estimated ðiÞ ðiÞ
tion gðY FK jf FK ; r2FK Þ using the parameters and compute
using Expectation–Maximization (E–M) algorithm. Dy(i,j) = YFK(i,j)  YPK(i,j).
iii. Complete step (ii) for all ensembles (1000) and for all spatial
3.4. Expectation–Maximization algorithm points for a given resolution.
iv. Signal to Noise Ratio (SNR) is computed as; The Difference of
EM algorithm is an iterative method and consists of two steps; future and past climate expected changes/standard devia-
the Expectation (E) step and the Maximization (M) step. tion (Dy). The expected weighted mean changes are
It begins with the initial guess of parameter h(0) for the param- obtained from Eqs. (6) and (7). This measures SNR for local
eter vector h = {wk, r2PK , r2FK }. In the E step, a quantity Zkst estimated climate change. For further details refer to Duan and
given present values of parameters; however Zkst takes two values; Philips (2010).
either 0 or 1. Where Zkst = 1 if ensemble member k is best forecast
for particular (s, t). Otherwise Zkst = 0.
4. Results and discussion
Thus, the E-step yields,
n o A linear regression based statistical downscaling method is
ðiÞ ðiÞ ðiÞ ðiÞ ðiÞ ðiÞ ðiÞ ðiÞ
Z kst ¼ gðY TP jhPK Þ  gðY TF jhPK Þ= Rk gðY TP jhPK Þ  gðY TF jhFK Þ ð4Þ
adopted for rainfall projections of ISMR at multiple resolutions,
viz., 0.05°, 0.25° and 0.50°, at monthly time scale to understand
While the M-step leads to, their projections and inter-compare them for potential differences.
X ðiÞ The results are presented in the following subsections which
ðiÞ
wk ¼ Z ks =S ð5Þ includes a comparison of simulations with observations for the his-
s
torical period, comparisons of PDF, comparison between BMA and
K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177 3169

(A) (B)

4 3

(C) (D)

4 3

(E) (F)

4 5

(G)

Fig. 3. The cumulative percentage of variance explained by NCEP/NCAR principal components (PCs) in the different zones for the zonal wind (u) at 500 hpa (90% variability
shown along with principal component number).

MMA, uncertainty across space and time, and the Signal to Noise GCMs, the absolute percentage errors in the mean and the standard
Ratio (SNR). deviation. The results show that characteristics of the mean are
quite faithful to the observational estimates. The spatial pattern
4.1. Comparison of results between the model and observations for of Indian monsoon rainfall is governed by orography due to the
historical period hilly regions of Himalayas, Western Ghats, Sapura, Vindhya and
Aravali ranges (Salvi et al., 2013; Rajendran and Kitoh, 2008). GCMs
The results of downscaled product are evaluated with respect to being coarse resolution models, cannot capture the spatial variabil-
observed rainfall data (Fig. 3). Figs. 3 and 4 show the grid-wise ity of ISMR. Here, we find that the statistical downscaling captures
details of the mean and standard deviation and are shown in the the occurrences of high rainfall in the windward side of the Wes-
order of the observed rainfall, the multi-model average from 19 tern Ghats and low rainfall for the leeward side. The Rajasthan
3170 K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177

Observed Mean Projected Ensemble Absolute Percentage


at 0.050 mean at 0.050 Errors in mean at 0.050

(A) (B) (C)

Observed Mean Projected Ensemble mean Absolute Percentage Errors


at 0.25 at 0.25 in mean at 0.25

(D) (E) (F)

Observed Mean Projected Ensemble mean Absolute Percentage Errors


at 0.50 at 0.50 in mean at 0.50

(G) (H) (I)

mm/day
Fig. 4. The absolute percentage errors in the mean of multi-model averaged downscaled simulations at 0.05°, 0.25° and 0.50° spatial resolutions, respectively.

receives scanty rainfall because the orientation of Aravali ranges, model. In the present analysis, for developing the statistical
and is also observed in the downscaling simulations. The North- downscaling, we have used the PCs in each of the seven zones
east part of India receives heavy rainfall because presence of separately and established the regression relation. We have
mountainous regions and is captured by statistical downscaling. shown in Fig. 5, the cumulative percentage variance explained
For majority of the grids over the Indian landmass, the abso- by the NCEP/NCAR PCs in different zones for the zonal wind (u)
lute percentage errors in the mean (Fig. 3(C), (F), and (I)) and at 500 hPa at 90% variability.
the standard deviation (Fig. 4(C), (F) and (I)) are found to be The linear regression in statistical downscaling assumes that
around 10%. Similarly the absolute percentage error for standard the residuals follow normal distribution. We check the normality
deviation is also less than 10% for most of the grids. However, assumptions for the residuals at multiple resolutions (Fig. 6). We
parts of the western and southern regions receive very scanty find that the errors at majority of the nodes follow normal distribu-
rainfall and traditionally, the statistical downscaling model can tion. We find that for 0.05°, 0.25° and 0.5° resolutions, around 15%,
only poorly capture this variability of rainfall (Salvi et al., 2013). 27% and 35% nodes violate the normality assumptions, respec-
In those regions, the absolute percentage error for standard devi- tively. The assumption is violated in the low rainfall regions which
ation is over 30%. The monthly means are expressed as mm/day. are mostly located in the western part and Jammu Kashmir regions
Along with the mean, the spatial pattern and temporal variability of India. This is considered as the limitation of this study. The use
of rainfall is also well captured by the statistical model with a of generalized linear regression or robust regression approaches
high degree of accuracy. The spatial dependence of rainfall may yield better results. The 0.5 degree downscaling has a slightly
between rainfall stations has also been well modelled by the higher percentage of around 35% violations.
K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177 3171

Observed Standard Projected Ensemble Standard Absolute Percentage Errors in


Deviaon at 0.050 deviaon at 0.050 Standard Deviaon at 0.050

(A) (B) (C)

Observed Standard Projected Ensemble Standard Absolute Percentage Errors in


Deviaon at 0.25 0 deviaon at 0.250 Standard Deviaon at 0.250

(D) (E) (F)

Observed Standard Projected Ensemble Standard Absolute Percentage Errors in


Deviaon at 0.5 0 deviaon at 0.50 Standard deviaon at 0.50

(G) (H) (I)

Fig. 5. The absolute percentage errors in the standard deviation of multi model averaged downscaled simulations at 0.05°, 0.25°and 0.50° spatial resolutions, respectively.

0.05 0 0.25 0 0.50 0


(A) (B) (C)

15% 27% 35%

Non-Normal Normal
Distribuon Distribuon

Fig. 6. Grids, where the residuals follow a normal distribution at 0.05°, 0.25° and 0.50° spatial resolutions, respectively. Fractions of grids, where residuals do not follow a
normal distribution, are also shown.

4.2. Skill score based PDF comparison which is based on the similarity of the PDFs computed with both
observations and simulations for the historic period (Fig. 7).
The present statistical downscaling method captures the mean Fig. 7a shows that the skill score obtained from the transfer func-
conditions very well for multiple resolutions over the historical tion based downscaling at different resolutions are almost similar;
period. Here, we also use the Skill score (S-score) for evaluation, however, such a conclusion is not valid for the interpolation
3172 K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177

PDF of Skill Score for 0.050 resoluon


PDF of Skill Score for 0.250 resoluon (A) PDF of Skill score for 0.050 Resoluon (B)
PDF of Skill score for 0.250 Resoluon
PDF of Skill Score for 0.50 resoluon further regridded to 0.05 0
PDF of Skill score for 0.500 Resoluon
further regridded to 0.050

Fig. 7. Skill scores for downscaled projections at three resolutions, viz., 0.05°, 0.25° and 0.50°. The spatial distribution is represented as a PDF. Panel A represents the skill
score at different resolutions. Panel B shows the skill score at 0.05°, where the downscaled simulations of 0.25° and 0.50° are further interpolated.

0.05 0 0.25 0 0.50 0

(A) (B) (C)

Fig. 8. Weights obtained for different GCMs at different resolutions with Bayesian modelling for 2080s.

method (Fig. 7b). We also perform downscaling at different resolu- are presented in Fig. 8 at different resolutions for the 2080s. We
tions, viz., 0.50°, 0.25° and 0.05° and then re-grid the first two into obtain consistently higher weights for GISS, CMCC, MIROC and
the latter one. We observe that model performance is best for 0.05° MPI at all the resolutions. Similarly, consistently lower weights
resolution and reduces with the coarseness of downscaling resolu- are assigned to BCC, BNU and CCCMA.
tion (Fig. 7b). The projections obtained with both MMA (Fig. 9a–c) and BMA
We apply a transfer function based statistical downscaling (Fig. 9d–f) show spatially non-uniform projected changes of ISMR
approach, which relies on the statistical relationship between the for the 21st century. Increases in precipitation are projected for
large scale circulation pattern (predictor) and local rainfall (pre- the Himalayan foothills, the west coast, and the northern regions,
dictand). The other approach of downscaling is a change factor while decreasing trends are observed in the Gangetic West Bengal
based method, which involves interpolation and bias correction. and regions of Central India with mixed results in Southern India.
Here, we regrid 0.5° and 0.25° datasets to 0.05° to check if such The increase/decrease in rainfall for future time slices is consistent
an interpolation method is really valid for the spatially varying with other studies with fine resolution models (Salvi et al., 2013;
monsoon rainfall. The results show, that interpolation method fails Shashikanth et al., 2013; Rupa Kumar et al., 2006; Rajendran and
to attain the skill which is possible to achieve with the transfer Kitoh, 2008). The orographic rainfall is expected to increase in west
function based method. This also points to a very important con- coast and the northeast part of India in the future in those studies
clusion on the possible failure of methods such as the Bias Correc- also. We find spatially non-uniform projected changes of ISMR
tion Spatial Disaggregation (BCSD), which are widely used and which is not visible in the original simulations from GCMs
based on interpolation and bias correction. The reason is probably (Shashikanth et al., 2013). Rajendran and Kitoh (2008) suggested
that the high spatial variability of the Indian monsoon rainfall. This that there will be changes in the moisture holding capacity of
is quite expected since the reduction of model performance is pri- atmosphere in the future; mean changes in the wind and conver-
marily due to interpolation without additional information from gence of moisture at 850 hPa, evaporation, and precipitable water
synoptic scale circulation. This also highlights a possible limitation (PWAT) are some of the causes for future changes in mean rainfall
of widely used change factor based statistical downscaling meth- over India. In general, we expect intensification of land–ocean tem-
ods, such as the BCSD (Wood et al., 2004), which is just a combina- perature gradient, a strengthening of the of low pressure trough,
tion of interpolation and bias correction. and more moisture in atmosphere (as per the Clausius–Clapeyron
relation), which may intensify the ISMR in the future (Kripalani
4.3. Comparison of results based on BMA and MMA et al., 2007). However, uneven heating may result in uneven
distribution of rainfall resulting non-uniform changes and the past
To know the plausible changes in future rainfall, we consider decades have shown a decreasing trend in ISMR despite the rapid
downscaled rainfall projections for RCP4.5 scenarios at 0.50°, warming of the Indian Ocean and the models are unable to capture
0.25° and 0.05° resolutions. We consider the projections obtained this trend (Saha et al., 2013). For the present study, we have
with both MMA and BMA. The weights obtained by using BMA considered a blend of predictors which increases our confidence
K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177 3173

Changes in the mean for Changes in the mean for Changes in the mean for
0.050 Resoluon 0.250 Resoluon 0.500 Resoluon
(A) (B) (C)
MMA MMA MMA

mm/day

(D) (E) (F)


BMA BMA BMA

(G) (H) (I)


MMA-BMA MMA-BMA MMA-BMA %

Fig. 9. Projected changes in mean rainfall for a future period (2070–2099) with respect to the base line period (1970–2005),with Multi-Model Average (MMA) at 0.05° (A),
0.25° (B) and 0.50° (C) and with BMA (D–F, respectively). The percentage difference between MMA and BMA are presented in panels G–I.

0.05 0 0.25 0 0.50 0


(A) (B) (C)

25% 28% 27%

Non-Normal Normal
Distribuon Distribuon

Fig. 10. Grids, where the projected changes in rainfall follow normal distribution at 0.05°, 0.25°and 0.50° spatial resolutions, respectively. Fractions of grids, where changes
do not follow normal distribution, are also shown.

in climate projections for the future. The statistical downscaling BMA, the changes in rainfall should follow normal distribution.
assumes stationarity in the relation between predictors and pre- Here we find (Fig. 10) 25%, 28%, 27% of the grids at resolutions
dictands. This is the main limitation of the present work. 0.05°, 0.25° and 0.5°, respectively, violate this assumption. The
We find that the changes projected with MMA are slightly more BMA should be applied for those grids with a modified maximum
on the positive side, compared to BMA. This suggests the possibil- likelihood function. This is considered as the potential scope for
ity of the existence of models with low weights projecting high future research.
positive changes. The differences between BMA and MMA are not The climate projections at global/regional scale are subject to
sensitive to the resolution of downscaling. For the application of uncertainty associated with the use of multiple models. Here we
3174 K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177

Uncertainty Changes in uncertaines measured with respect to 2020s


2020s 2050s 2080s
(A) (B) (C)

(D) (E) (F)

(G) (H) (I)

Fig. 11. Multi-model uncertainty during 2020s and its changes with time. The uncertainty is measured as coefficient of variation.

use the coefficient of variation (CV) (Shashikanth et al., 2013; resolutions. This is in agreement with Hawkins and Sutton
Woldemeskel et al., 2012) as a measure of uncertainty for ISMR (2009), where the prominence of climate change signal is argued
projections at different resolutions and for different time windows to emerge at longer lead times. We also find that SNR does not
2020s (2010–2039), 2050s (2040–2069) and 2080s (2070–2099). depend on the resolution of downscaling. Statistical downscaling
The relative changes in uncertainties across different resolutions uses the statistical relationship between the fine resolution rainfall
are measured with respect to 2020s. We find that the model uncer- and coarse resolution circulation patterns and hence this is mostly
tainty is highest for the 2020s and then it gradually decreases with affected by the uncertainty in the circulation patterns due to the
the lead time (Fig. 11). This is true at all resolutions. The possible use of multiple GCMs. Increase in resolution does not add any
reason is that with the increase in lead time, the climate change source of uncertainty into the downscaling method and hence
signal becomes more prominent compared to internal variability the SNR and uncertainty remain unaffected by the change in reso-
and model uncertainty due to the imposed radiative forcings lution. Though fine resolution modelling improves the simulation
(Hawkins and Sutton, 2009), resulting in a decrease in the CV of local scale climatology (Fig. 7), it does not necessarily add any
values. value in terms of changes (Fig. 9), uncertainty (Fig. 11) and SNR
(Fig. 12). The primary purpose for downscaling in climate change
4.4. Estimation of SNR at multiple resolutions studies is to model the projected changes and here we do not find
any significant improvements or differences in the projected
The SNR in climate change projections indicates the strength of changes. However, the selection of optimal resolution for down-
the climate signal with respect to the uncertainty associated with scaling is essential for specifically capturing the orography and
the use of multiple models. Here we compute SNR for the three region-specific land, ocean, atmosphere interactions which are
resolutions (0.05°, 0.25°, and 0.50°) for different time periods generally not considered in coarse resolution GCM simulations.
(2020s, 2050s and 2080s) (Fig. 12). With the reduction of model The present analysis points to the need for research activities in
uncertainty, the SNR increases from the 2020s to 2080s for all understanding the optimum resolution required in downscaling
K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177 3175

0.05 0 0.25 0 0.50 0

S (A) (B) (C)


N
R
2
0
2
0

(D) (E) (F)


S
N
R
2
0
5
0

(G) (H) (I)


S
N
R
2
0
8
0

Fig. 12. The Signal to Noise Ratio (SNR) for projections at multiple resolutions for different time periods. The pie chart shows the fraction of grid points having SNR more/less
than 0.5.

(specific to a region), to capture the orographic and other local uncertainty in statistical downscaling. This is specifically
effects, and at the same time, not to increase the resolution unless because the key uncertainty in statistical downscaling is due
beneficial effects are expected on the SNR to justify the high com- to the uncertainty in synoptic scale circulation pattern simu-
putational efforts. lated by the parent models or the reanalysis data employed in
developing the statistical relations (Kannan et al., 2014). As
5. Conclusions the observed rainfall data at multiple resolutions are obtained
from the same agency, rainfall data does not add resolution-
In this study, we use statistical downscaling for projections of dependent uncertainty. However, the contrary is not necessar-
ISMR at multiple resolutions. The primary objective is to under- ily true; i.e., increasing resolutions does not necessarily
stand the improvements obtained by increasing the spatial resolu- guarantee the usability or the usefulness of the downscaled
tion from 0.50° to 0.05°. Following are the main conclusions from product at arbitrarily high resolutions. Caution is thus needed
the present study. in using other metrics for determining optimal resolutions for
statistical downscaling. This is reemphasized below in the con-
1. The skill score plots (Fig. 7) show improvements in simulating text of SNR.
climatology at the finest resolution (0.05°) with respect to 3. We also find that the SNR of climate change projections does
observed data, when we use direct statistical downscaling at that not improve by increasing resolution. Here SNR denotes the
finest resolution in comparison to that of simulations (at 0.05°) strength of the signal of climate change with respect to the
interpolated from coarser resolution downscaled data (0.5° and uncertainty. As there is no significant change in uncertainty
0.25°). This result also highlights the failure of the interpolation with the enhanced resolution, the SNR also remains the same.
method, which are often coupled with a bias correction (Wood This highlights that a finer resolution in statistical downscaling
et al., 2004), without any specific regional modelling. does not improve the reliability of projected changes; however
2. Enhanced resolution increases the computational requirements an optimal resolution for downscaling is necessary to capture
for any downscaling and hence it is necessary to evaluate the the major local and orographic factors and other region-specific
finest resolution model for payoff in terms of uncertainty and convective processes related to the land–atmosphere–ocean
SNR. We observe that enhanced resolution does not affect the interactions.
3176 K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177

4. The present study shows that there is no significant difference Ghosh, S., Mujumdar, P.P., 2006. Future rainfall scenario over Orissa with GCM
projections by statistical downscaling. Current Sci. India 90 (3), 396–404.
between the BMA and MMA projections at any resolution;
Ghosh, S., Mujumdar, P.P., 2009. Climate change impact assessment- uncertainty
though, the projected changes with MMA are slightly more on modeling with imprecise probability. J. Geophys. Res. Atmos. 114, D18113.
the positive side, compared to BMA. This is because of the use http://dx.doi.org/10.1029/2008JD011648.
of a relatively large number of models. With the increase in Giorgi, F., Mearns, L.O., 2002. Calculation of average, uncertainty range, and
reliability of regional climate changes from AOGCM simulations via the
the number of models, the individual weights are assigned to reliability ensemble averaging REA method. J. Clim. 15, 1141–1158.
lower values. For example, if there are 5 models, with MMA, Groppelli, B., Bocchiola, D., Rosso, R., 2011. Spatial downscaling of precipitation
the individual weights will be 0.2; but with 20 models, the from GCMs for climate change projections using random cascades: a case study
in Italy. Water Resour. Res. http://dx.doi.org/10.1029/2010WR009437.
weights are reduced to 0.05. The change in the weights, where Gutmann et al., 2014. An intercomparison of statistical downscaling methods used
they themselves are very low, because of the use of a ‘sufficient’ for water resource assessments in the United States accepted article. Water Res.
number of models, may not result in significant changes in the Res. http://dx.doi.org/10.1002/2014/WR015559.
Hawkins, Ed., Sutton, Rowan., 2009. The potential to narrow uncertainty in regional
ultimate multi-model projections. climate predictions. Bull. Am. Meteorol. Soc. 90, 1095–1107. http://dx.doi.org/
10.1175/2009BAMS2607.1.
Further work is needed in determining what this ‘sufficient’ Hawkins, Ed., Sutton, Rowan., 2011. The potential to narrow uncertainty in regional
climate predictions. Clim. Dyn. 37, 407–418. http://dx.doi.org/10.1007/s00382-
number is and what the optimal resolution for statistical 010-0810-6.
downscaling would be as the resolution of the regional and global Hewitson, B.C., Crane, R.G., 1992. Large-scale atmospheric controls on local
GCMs and the available data continue to increase. The real ultimate precipitation in tropical Mexico. Geophys. Res. Lett. http://dx.doi.org/10.1029/
92GL01423.
decider will of course be the use and usability of the products in
IPCC, 2007. Climate change 2007: the physical science basis contribution of working
decision-making and policy frameworks. The exercise here is but Group I to the fourth assessment report of the intergovernmental panel on
the first step in exploring these requirements for decision-support. climate change. In: Solomon, S., Qin, D., Manning, M., Chen, Z., Marquis, M.,
Averyt, K.B., Tignor, M., Miller, H.L. (Eds.), Cambridge University Press,
Cambridge, UK/New York.
Acknowledgements Kalnay, N.Y. et al., 1996. The NCEP/NCAR 40-years reanalysis project. Bull. Am.
Meteorol. Soc. 773, 437471.
Kannan, S., Ghosh, S., 2013. A nonparametric kernel regression model for
We acknowledge the World Climate Research Programme’s downscaling multisite daily precipitation in the Mahanadi basin. Water
working Group on coupled Modelling, which is responsible for Resour. Res. 49. http://dx.doi.org/10.1002/wrcr.20118.
Kannan, S., Ghosh, S., Mishra, V., Salvi, K., 2014. Uncertainty resulting from multiple
CMIP, and we thank the modelling groups for producing and mak- data usage in statistical downscaling. Geophys. Res. Lett. http://dx.doi.org/
ing available their model output. For CMIP the U.S. Department of 10.1002/2014GL060089.
Energy’s Program for Climate Model Diagnosis and Intercomparison Khalili, M., Brissette, F., Leconte, R., 2009. Stochastic multi-site generation of daily
weather data. Stochastic Environ. Res. Risk Assess. 23, 837–849. http://
(PCMDI) provides coordinating support and led development of
dx.doi.org/10.1007/s00477-008-0275-x.
software infrastructure in partnership with the Global Organization King, L., Irwin, S., Sarwar, R., McLeod, A.I., Simonovic, S.P., 2012. The effects of
for Earth System Science Portals. We would like to thank APHRO- climate change on extreme precipitation events in the Upper Thames River
Basin: a comparison of downscaling approaches. Can. Water Res. J. 373, 253–
DITE, Japan for making available 0.050, 0.250 and 0.500 resolution
274.
observed data. We particularly thank Dr. Kamaguchi of APHRODITE, Knutti, R., 2010. The end of model democracy. Clim. Change 102, 395–404. http://
for making available the 0.05° observed data. Raghu Murtugudde dx.doi.org/10.1007/s10584-010-9800-2.
gratefully acknowledges the National Monsoon Mission funding Kripalani, H.R., Oh, J.H., Kulkarni, A., Sabade, S.S., Chaudari, H.S., 2007. South Asian
summer monsoon precipitation variability: coupled climate model simulations
for partial support. K Shashikanth acknowledges Qingyun Duan and projections under IPCC AR4. Theor. Appl. Climatol. http://dx.doi.org/
for his help in understanding Bayesian Multimodel Averaging. 10.1007/s0074-006-282-0.
Liang et al., 2012. Regional climate weather research and forecasting model. Bull.
Am. Meteorol. Soc. http://dx.doi.org/10.1175/BAMS-D-11-00180.1.
References Maraun, D. et al., 2010. Precipitation downscaling under climate change: recent
developments to bridge the gap between dynamical models and the end user.
Rev. Geophys. 48, RG3003. http://dx.doi.org/10.1029/2009RG000314.
Anandhi, A., Srinivas, V.V., Nanjundiah, R.S., Kumar, D.N., 2008. Downscaling
Meherotra, R., Sharma, A., 2005. A non parametric non-homogeneous hidden
precipitation to river basin for IPCC SRES scenarios using support vector
Markov model for downscaling of multisite daily rainfall occurrences. J.
machines. Int. J. Climatol. 28 (3), 401–420. http://dx.doi.org/10.1002/joc.1529.
Geophys. Res. 110, D16108. http://dx.doi.org/10.1029/2004/JD005677.
Bardaossy, A., Plate, E.J., 1992. Space-time model for daily rainfall using
Meherotra, R., Sharma, A., 2006. A non parametric stochastic downscaling frame
atmospheric circulation patterns. Water Resour. Res. 28 (5), 1247–1259.
work for daily rainfall at multiple locations. J. Geophys. Res. 111, D15101.
http://dx.doi.org/10.1029/91WR02589.
http://dx.doi.org/10.1029/2005JD006637.
Convoy, D., Jones, P.D., 1998. The use of weather types and air flow indices for GCM
Mehrotra, R., Sharma, A., 2010. Development and application of a multisite rainfall
downscaling. J. Hydrol. 212–213, 348–361.
stochastic downscaling framework for climate change impact assessment.
Corte-Real, J., Zhang, X., Wang, X., 1995. Downscaling GCM information to regional
Water Resour. Res. 46, W07526. http://dx.doi.org/10.1029/2009WR008423.
scales: a non-parametric multivariate regression approach. Clim. Dyn. 11,
Menon, A., Levermann, A., Schewe, J., 2013. Enhanced future variability during
413–424.
India’s rainy season. Geophys. Res. Lett. 40, 3242–3247. http://dx.doi.org/
Deser, C., Knutti, R., Solomon, S., Phillips, A.S., 2012. Communication of the role of
10.1002/grl.50583.
natural variability in future North American climate. Nat. Clim. Change 2,
Miao, Chiyuan, Duan, Qingyun, Sun, Qiaohong, Li, Jianduo, 2013. Evaluation and
775–779. http://dx.doi.org/10.1038/nclimate1562.
application of Bayesian multi-model estimation in temperature simulations.
Dibike, Y.B., Coulibaly, P., 2005. Hydrologic impact of climate change in the
Prog. Phys. Geogr. 1–18. http://dx.doi.org/10.1177/030913313494961.
Saguenay watershed: comparison of downscaling methods and hydrologic
Miao, Chiyuan, Duan, Qingyun, Sun, Qiaohong, Huang, Yong, Kong, Dongxian, Yang,
models. J. Hydrol. 307, 144–163.
Tiantian, Ye, Aizhong, Di, Zhenhua, Gong, Wei, 2014. Assessment of CMIP5
Dobler, A., Ahrens, B., 2011. Four climate change scenarios for the Indian summer
climate models and projected temperature changes over Northern Eurasia.
monsoon by the regional climate model COSMO-CLM. J. Geophys. Res. 116,
Environ. Res. Lett. http://dx.doi.org/10.1088/1748-9326/9/055007.
D24104. http://dx.doi.org/10.1029/2011JD016329.
Parthasarathy, B., Munot, A.A., Kothawale, D.R., 1996. Monthly and Seasonal Rainfall
Duan, Qingyun, Philips, Thomas J., 2010. Bayesian estimation of local signal and
Series for All-India Homogeneous Regions and Meteorological Subdivisions.
noise in multimodel simulations of climate change. J. Geophys. Res. 115,
Indian Institute of Tropical Meteorology, Pune, India, pp. 1871–1994.
D18123. http://dx.doi.org/10.1029/2009/JD013654.
Perkins, S.E., Pitman, A.J., Holbrook, N.J., McAneney, J., 2007. Evaluation of the AR4
Fowler, H.J., Blenkinsop, S., Tebaldi, C., 2007. Linking climate change modelling to
climate models simulated daily maximum temperature, minimum temperature
impacts studies: recent advances in downscaling techniques for hydrological
and precipitation over Australia using probability density functions. J. Clim. 20,
modelling. Int. J. Climatol. 27, 1547–1578.
4356. http://dx.doi.org/10.1175/JCLI4253.1.
Ghosh, S., 2010. SVM-PGSL coupled approach for statistical downscaling to predict
Prudhomme, C., Davies, H., 2009. Assessing uncertainties in climate change impact
rainfall from GCM output. J. Geophys. Res. 115, D22102. http://dx.doi.org/
analyses on the river flow regimes in the UK. Part 2: Future climate. Clim.
10.1029/2009JD013548.
Change 93, 197–222.
Ghosh, S., Katkar, S., 2012. Modeling uncertainty resulting from multiple
Prudhomme, C., Reynard, N., Crooks, S., 2002. Downscaling of global climate models
downscaling methods in assessing hydrological impacts of climate change.
for flood frequency analysis: where are we now? Hydrol. Process. 16, 1137–
Water Resour. Manage. 26, 3559–3579. http://dx.doi.org/10.1007/s11269-012-
1150.
0090-5.
K. Shashikanth et al. / Journal of Hydrology 519 (2014) 3163–3177 3177

Raftery, A.E., Gneiting, T., Balabdaoui, F., Polakowski, M., 2005. Using Bayesian Von Storch, H., 1999. On the use of ‘‘inflation’’ in downscaling. J. Clim. 12, 3505–
model averaging to calibrate forecast ensembles. Mon. Weather Rev. 133, 1155– 3506.
1174. http://dx.doi.org/10.1175/MWR2906.1. Watanabe, S., Kanae, S., Seto, S., Yeh, P.J.-F., Hirabayashi, Y., Oki, T., 2012.
Räisänen, J., 2007. Objective comparison of patterns of CO2 induced climate change Intercomparison of bias-correction methods for monthly temperature and
in coupled GCM experiments. Clim. Dyn. 13, 197–211. precipitation simulated by multiple climate models. J. Geophys. Res. 117,
Raje, D., Mujumdar, P.P., 2009. A conditional Random field based downscaling D23114. http://dx.doi.org/10.1029/2012JD018192.
method for assessment of climate change impact on multisite daily Wayne, G.P., 2013. Beginners guide to representative concentration pathways.
precipitation in the Mahanadi basin. Water Resour. Res. 45, W10404, Skeptical Sci. version 1.
doi:1029/2008WR007487. Webster, P.J., Magana, V.O., Palmer, T.N., Shukla, J., Tomas, R.A., Yanai, M., Yasunari,
Rajendran, K., Kitoh, A., 2008. Indian summer monsoon in future climate projection T., 1998. Monsoons: processes, predictability and the prospects for prediction. J.
by a super high resolution global model. Curr. Sci. 95 (11), 1560–1569. Geophys. Res. 103, 14451–14510.
Rupa Kumar, K.R., Sahai, A.K., Kumar, K.K., Patwardhan, S.K., Mishra, P.K., Revadekar, Wilby, R.L., Dawson, C.W., 2004. Using SDSM version 3.1—a decision support tool for
J.V., Kamala, K., 2006. High-resolution climate change scenarios for India for the the assessment of regional climate change impacts. User Manual.
21st century. Curr. Sci. 90 (3). Wilby, R.L., Wigley, T.M.L., 1997. Downscaling general circulation model output: a
Saha, S.K., Pokhrel, S., Chaudhari, H.S., Dhakate, A., Shewlale, S., Sabeerali, C.T., review of methods and limitations. Prog. Phys. Geogr. 21, 530–548.
Salunke, K., Hazra, A., 2013. Improved Simulation of Indian summer monsoon in Wilby, R., et al., 2004. Guidelines for use of climate scenarios developed from
latest NCEP climate forecast system free run, Int. J. Climatol. http://dx.doi.org/ statistical downscaling methods. <http://www.narccap.ucar.edu/doc/tgica-
10.1002/joc.3791. guidance-2004.pdf> (accessed 719 10.08.13).
Salvi, K., Ghosh, S., Kannan, S., 2013. High-resolution multisite daily rainfall Wilks, D.S., 1999a. Multisite downscaling of daily precipitation with a stochastic
projections in India with statistical downscaling for climate change impacts weather generator. Clim. Res. 11, 125–136.
assessment. J. Geophys. Res. Atmos. 118. http://dx.doi.org/10.1002/jgrd.50280. Wilks, D.S., 1999b. Multi-site downscaling of daily precipitation with a stochastic
Schiermeier, Q., 2010. The real holes in the climate science. Nature 463, 284–287. weather generator. Clim. Res. 11, 125–136.
Schoof, J.T., 2013. Statistical downscaling in climatology. Geogr. Compass 7 (4), Wise, M.A., Calvin, K.V., Thomson, A.M., Clarke, L.E., Bond-Lamberty, B., Sands, R.D.,
249–265. http://dx.doi.org/10.111/gec3.12036. Smith, S.J., Janetos, A.C., Edmonds, J.A., 2009. Implications of limiting CO2
Shashikanth, K., Kaustubh, S., Ghosh, S., Rajendran, K., 2013. Do CMIP5 simulations concentrations for land use and energy. Science 324, 1183–1186.
of Indian summer monsoon rainfall differ from those of CMIP3? Atmosp. Sci. Woldemeskel, F.M., Sharma, A., Sivakumar, B., Mehrotra, R., 2012. An error
Lett.. http://dx.doi.org/10.1002/asl2.466, 10.1002. estimation method for precipitation and temperature projections for future
Stainforth, D.A., Allen, M.R., Tredger, E.R., Smith, L.A., 2007. Confidence, uncertainty climates. J. Geophys. Res. 117, D22104. http://dx.doi.org/10.1029/
and decision-support relevance in climate predictions. Philos. Trans. R. Soc. 2012JD018062.
2007, 365. http://dx.doi.org/10.1098/rsta.2007.2074. Wood, A.W., Leung, l.R., Sridhar, V., Lettenmaier, D.P., 2004. Hydrologic implications
Taylor, Karl E. et al., 2012. An overview of CMIP5 and the experiment design. Bull. of dynamical and statistical approaches to downscaling climate model outputs.
Am. Meteorol. Soc. 93, 485–498. http://dx.doi.org/10.1175/BAMS-D-11- Clim. Change 62, 189–216.
00094.1. Yatagai, A. et al., 2012. APHRODITE: constructing a long-term daily gridded
Tebaldi, C., Mearns, L.O., Nychkam, D., Smith, R.W., 2004. Regional probabilities of precipitation dataset for Asia based on a dense network of rain gauges. Bull.
precipitation change: A Bayesian analysis of multi-model simulations. Geophys. Am. Meteor. Soc. 939 (1401–1415), 727. http://dx.doi.org/10.1175/BAMS-D-11-
Res. Lett. 31, L24213, doi:L10.1029/2004GL021276. 00122.1.
Tebaldi, C., Mearns, L.O., Nychkam, R.W., Smith, D., 2005. Quantifying uncertainty in Yip, Stan, Christopher, A.T., Ferro, David B., Stephenson, Ed Hawkins, 2011. A simple,
projections of regional climate change: a Bayesian approach to the analysis of coherent framework for partitioning uncertainty in climate predictions. J. Clim.
multimodel ensembles. J. Clim. 18, 1524–1540. 24, 4634–4643. http://dx.doi.org/10.1175/2011JCLI4085.1.
Von Storch, H., 1995. Inconsistencies at the interface of climate impact studies and
global climate research. Meteorol. Zeitschrift (4NF), 72–80.

You might also like